noticias radar ia agentes de ia ciberseguridad gemini

Wednesday's radar was about monitoring agents; Thursday's is about what they cost. In twenty-four hours, Google has priced its fast model at $0.75 per million tokens with an expiration date, and Anthropic has debuted Claude Fable 5.1 without touching the rate but cutting cache costs by 75%, while both labs and Google lock their vulnerability-finding models behind a guest-list door. And the money confirms yesterday's thread: Palo Alto Networks is paying 500 million for a two-year-old agent startup, and HiddenLayer is raising 100 to monitor third-party agents.

Google lowers Gemini 3.8 Flash to $0.75 per million tokens, but the price expires on December 31st

Google launched Gemini 3.8 Flash yesterday, its fourth Flash model in four months, with a launch price of $0.75 per million input tokens and $3.75 for output that expires on December 31, 2026: starting in January, it moves to 1.50 and 7.50, exactly double. It takes in a million tokens and outputs 64,000, and Google claims that in long, document-heavy workflows, it completes more than three times the tasks of Gemini 3.7 Flash; in the DeepSWE v1.1 software engineering benchmark, it outperforms most larger top-tier models. The nuance missing from Google's note was provided by Artificial Analysis: the model rises from 56 to 59 on its index, tying with GPT-5.6 Sol and Grok 4.6, but each task costs $0.58, 40% more than 3.7 Flash, because it reasons through more steps and calls tools more frequently; even so, it is about six times cheaper per task than Claude Fable 5.1. For your company: the price per token no longer tells you what you're going to pay, and this is the perfect example. Measure the cost per completed task in your own workflows—a classified invoice, an answered email, a finalized report—before switching models, because the cheap model that "thinks more" may cost you more per result than the one it replaces. And note the date on the budget sheet: everything you plan for 2027 with this model goes at 1.50 and 7.50, not 0.75; a launch rate is a promotion, not a price. Source

Anthropic debuts Claude Fable 5.1 without raising the rate and lets your logs stay in your cloud

Anthropic released two twin versions of its most advanced model on Tuesday: Claude Fable 5.1, open in its API and on Amazon Bedrock, Google Cloud, and Microsoft Azure, and Claude Mythos 5.1, the same model with more permissive safeguards and reserved for verified cybersecurity and life sciences organizations. The rate doesn't change—$10 per million input tokens and $50 for output—but cache reads drop from 1 to $0.25, a 75% reduction, which the company claims lowers a typical workload by 25% and a highly agentic one by up to 45%. The same day, it announced Enterprise Frontier Safeguards, the response to a specific complaint: with Fable 5, it had moved from zero retention to saving thirty days of data to detect abuse spread across sessions and accounts, and regulated customers pushed back, led by security heads from Goldman Sachs, Morgan Stanley, Citi, Bank of America, and Wells Fargo. The solution: activity logs live in the customer's own storage (S3, Azure Blob, or Google Cloud Storage) under their keys, automatic abuse detection runs there, and alerts go to the customer, not to an Anthropic human reviewer; it's free (you only pay for cloud storage), was designed with over a hundred customers, and will roll out in phases this fall, with provisional zero retention for those who meet requirements. For your company: two things. First, the 45% savings doesn't come by default; it comes from design: cache only pays off if your agent reuses a long and stable context—instructions, manuals, house policy—placed always at the beginning and without changes; if every call rewrites the prompt, you pay the full rate. Second, the objection "my data goes to a third-party server" just received a response in contract format, and that format serves as a template for any provider, including the SaaS selling you "AI included": where do the logs live, under what keys, who reads them, and how many days are they kept. Four questions for the data protection form; if the provider can't answer them, that's your answer. Source

OpenAI and Google lock down their vulnerability hunters: your patching window is shortening

OpenAI said on Tuesday that Astra, its next model, is the first to reach the "critical" level of cybersecurity capability in its preparedness framework: perfect score on ExploitBench, the test measuring if a model converts a known vulnerability into a working attack; two unknown vulnerabilities discovered and exploited on its own in modified tests, and a browser sandbox escape to execute commands on the machine. It rejects 91.5% of attempts to bypass its filters on cyber topics compared to 59% for GPT-5.6 Sol, it will be out "soon," and its most advanced offensive capabilities will be limited to a group of testers and then to the Daybreak Blue program. A day later, Google presented Gemini 3.8 Flash Cyber: over 70% success rate in its internal vulnerability discovery benchmark across twenty programming languages, 2.6 times more correct patches for Chrome than the best commercial model its security team tested, and a critical core vulnerability found in less than two hours when the norm is months; access is only through the Fairwind program, for governments, critical infrastructure operators, and software maintainers. With Anthropic's Mythos 5.1, that's three labs in one week putting a guest list on the same capability. For your company: the door separates you from the model, not the attacker from the technique forever, so the working hypothesis is that the time between a flaw being published and someone exploiting it is measured in hours, not weeks. Three things that cost no money. A list of everything you have facing the internet with its version—firewalls, VPNs, ERP web access, WordPress plugins—and a patching deadline in days, not quarters, for that list and only that list. If a provider hosts it, ask in writing what their patching window is, because "they manage it" is not an answer. And nothing exposed without multi-factor authentication, which is what keeps working when the patch arrives late. Source

Palo Alto pays 500 million for a two-year-old agent and HiddenLayer raises 100: monitoring agents already has buyers

Palo Alto Networks has acquired Console for $500 million in cash and stock, according to TechCrunch sources; the deal was announced Tuesday without a price. Console is two years old, had raised 29 million—a 6.2 seed led by Thrive Capital and a 23 Series A with DST Global—at a 157 million valuation, and sells agents that solve internal IT support without a human: resetting passwords, granting access to Figma or Miro, fixing the routine. It goes into Cortex, Palo Alto's threat detection platform, so the security team can investigate and resolve alerts by speaking in natural language; CEO Nikesh Arora, who was also an angel investor in Console, says Cortex gains "arms and legs," and it is Palo Alto's seventh acquisition in 2026. The same day, HiddenLayer closed a 100 million Series B led by Delta-v Capital with Ten Eleven, Morgan Stanley, Microsoft's M12, and Booz Allen: it protects models and agents at runtime, just added a module to monitor coding agents while they work, and says its recurring revenue has grown more than tenfold in a year to tens of millions, 90% from new customers. Gartner puts a number on the market: $2.83 billion in AI security spending in 2026, up 83% from 2025, and 4.78 billion in 2027. And money keeps flowing into the application layer too: Wonderful, which sells an operating system for agents in 35 countries, has raised 550 million at a 5 billion valuation, more than double its March valuation, with Salesforce joining for the first time. For your company: the helpdesk agent you hired from a startup could be a module of a security giant in two years, with a different price sheet and a different roadmap; at this point, the change-of-control clause and the right to export your workflows and settings aren't paranoia, they're the standard. And the Gartner figure is a yardstick to measure yourself: if your AI budget has zero dollars and zero hours to govern and monitor what you deploy, you are below what the market already takes for granted. Three deals in the category in 48 hours—AIR, Console, HiddenLayer—and the first step is still free: an inventory of what agents you have running, what connectors they use, and what credentials they exit with. Source

What to watch tomorrow?

Bloomberg reports that Nvidia could sign the acquisition of Hugging Face this very week for 12.9 billion plus a retention package of about 1 billion for the staff, nearly 14 billion in total: it's the thread open since August 25, and the signing would change the ownership of the open model repository used by half the sector. And while models drop in price at the top, the bill at the bottom keeps rising: Broadcom reported 16.7 billion in AI chip revenue for the quarter, up 221%, and forecasts 21.7 billion for the next; today's cuts come from competition between models, not from compute getting cheaper.

Watch the 1-minute video